Good morning. It's Thursday, April 23, and we're covering AI drug discovery breakthroughs, Google’s next-gen TPU chips, the rise of collaborative AI agents, and more.

Don’t forget to check out the results of yesterday’s poll. ↓

OpenAI Introduces Workspace Agents
Rather than improving individual productivity, the new agents are designed to handle shared, cross-team workflows that typically stall on coordination and handoffs. ChatGPT workspace agents, powered by Codex, can run tasks like reporting, lead qualification, and vendor risk analysis in the cloud, continuing work asynchronously and integrating with tools like Slack. Available in research preview for Business, Enterprise, Edu, and Teachers plans, they will remain free until May 6, 2026, before shifting to credit-based pricing.

Google Expands Cloud Bet With Thinking Machines
Despite already partnering with NVIDIA, Thinking Machines Lab is turning to a rival’s cloud to power its most compute-intensive work. The startup, founded by Mira Murati, has signed a multibillion-dollar agreement with Google Cloud to access AI infrastructure built on NVIDIA’s GB300 GPUs for model training and deployment. The non-exclusive deal, valued in the single-digit billions, gives Thinking Machines early access to systems offering roughly 2× speed improvements.

Google Brings AI Overviews To Gmail
Instead of opening threads, workplace users can now get answers synthesized across multiple emails instantly. Gmail is adding AI Overviews—previously used in Google Search—to summarize inbox content and respond to natural-language queries about topics like projects, invoices, and travel. The feature rolls out following its earlier availability in consumer tiers like Google AI Pro and Ultra, and will now extend to Business, Enterprise, and Education Workspace plans.

Cursor Pushes Into Slack With PR Automation
Instead of starting work in a dedicated tool, developers can now trigger coding tasks directly from chat threads. Cursor announced a Slack integration that lets users mention the assistant to generate pull requests using context from messages and surrounding channels. The system streams updates in real time as it builds code changes, aiming to collapse the gap between discussion and execution.

Higgsfield is the world's first AI-powered streaming platform, dropping original series made entirely by artificial intelligence. No studios, no gatekeepers, no $200M budgets. Just stories that shouldn't exist yet, streaming right now for free.

Stream for Free.

Tencent and Alibaba are in talks to invest in Chinese AI startup DeepSeek, which is seeking to raise at least $300 million at a valuation exceeding $20 billion. The fundraising marks DeepSeek’s first external capital raise after operating largely as an open-source AI lab with minimal revenue.

The potential valuation reflects strong investor demand and comparisons to peers like Moonshot AI, despite ongoing questions about DeepSeek’s business model. → Continue reading here. (Paywall)

Startup 10x Science announced a $4.8 million seed round, aiming to solve a growing bottleneck in AI-driven drug discovery: analyzing and validating the surge of predicted drug candidates. Founded in December 2025 by former Stanford researchers, the company combines AI agents with mass spectrometry data to help scientists characterize complex molecules more efficiently.

While AI tools like protein-structure prediction models have accelerated candidate generation, downstream analysis remains slow and resource-intensive. 10x’s platform targets that gap, offering pharmaceutical companies a subscription-based tool to speed up research and reduce reliance on costly in-house infrastructure. → Read the full article here.

Google introduced its eighth-generation Tensor Processing Units—TPU 8t and TPU 8i—at Cloud Next, targeting the growing demands of “agentic” AI systems. TPU 8t is designed for large-scale model training, delivering nearly 3× the compute performance of its predecessor, while TPU 8i focuses on low-latency inference with up to 80% better performance-per-dollar.

The chips reflect a shift toward specialized infrastructure as AI models increasingly rely on multi-step reasoning and autonomous workflows. Both chips will power Google’s AI Hypercomputer and are expected to become generally available later in 2026. → Read the full article here.

Sony AI’s table tennis robot, Ace, defeated elite human players in three out of five matches played under official rules. The system demonstrated advanced capabilities, including handling spin, reacting to net deflections, and executing complex shots. Ace uses multiple cameras and AI models trained on 3,000 hours of simulation to track ball movement and make split-second decisions.

Sullivan & Cromwell Admits AI Filing Errors: Elite Wall Street firm apologizes after hallucinated citations and misquoted law appeared in April 9 bankruptcy filing, exposing gaps in AI oversight and legal review safeguards.

NVIDIA Joins $1B Vast Data Round: AI data infrastructure firm hits $30 billion valuation, tripling since 2023, as investors bet on systems powering millions of GPUs amid surging enterprise AI demand.

Claude Launches AI Code Review Agents: Anthropic previews “/ultrareview,” a cloud-based tool running multiple bug-hunting agents that auto-deliver findings to CLI or desktop, targeting critical code checks before deployment.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team